Your customer no longer always types “who installs warehouse camera analytics in Baku?” into Google. A growing share asks ChatGPT, Perplexity or Gemini instead — and gets back an answer naming three or four companies. If you are on that list, they contact you. If you are not, you effectively do not exist.
This field is called GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization). The trouble is that most of what is written about it is either an empty promise or old SEO advice under a new name. This article sticks to measured findings.
How AI search differs from SEO
Classic search hands you ten links and lets you choose. An AI system builds one answer and cites two or three sources inside it. The difference is not quantity — it is logic:
- Ranking tenth on Google still brings some traffic. In an AI answer you are either named or you are not.
- Google matches keywords; a language model looks for an answer. Your page has to state it plainly.
- Google ranks a site; the model lifts a sentence. Writing quotable sentences is a separate skill.
What the numbers say
Ranking on page one is not enough. Ahrefs analysed 15,000 queries and found that only 12% of the URLs cited by AI systems overlap with Google's top ten. The remaining 88% come from pages that do not rank on the first page.
Each system picks different sources. Only about 11% of the domains cited by ChatGPT and by Perplexity are the same. Visibility in one does not guarantee visibility in the other.
Not every answer involves a search. For roughly two thirds of queries ChatGPT does not go online at all — it answers from training data. Optimisation does not reach that path; only long-term brand recognition does.
A handful of giants take most citations. Reddit, Wikipedia, YouTube, LinkedIn and a few large media outlets absorb around two thirds of all citations.
That last figure should not discourage you — it points the way. In narrow, concrete and local queries those giants have no answer. Wikipedia will not explain door monitoring in a cold storage warehouse. You can.
Six steps that work
1. Structure content as questions and answers
A model reads text in fragments and wants the question and the answer close together. Use a specific question as a heading and put a short, complete answer directly underneath.
2. Add FAQPage structured data
This schema exposes question-answer pairs in a form machines read directly, serving both Google's FAQ block and model citations. One JSON-LD block per page is enough. Official documentation: Google Search Central — FAQPage.
3. Be concrete, not general
Models pick sentences that can be measured. “Our service is high quality” will never appear in an answer. “Ready scenarios for fourteen sectors, a one-week trial, no charge” can be quoted. The rule: every claim carries a number, a duration or a condition.
4. Show the author and the source
Models weigh trust signals: who wrote this, what experience it rests on, whether the company is real. Anonymous text is the weakest format. In practice: name and role under every article, address and contact details visible on the site.
5. Do not block the crawlers
If GPTBot, ClaudeBot, PerplexityBot or Google-Extended are blocked in your robots.txt, no amount of optimisation will help. This is sometimes deliberate and sometimes accidental — worth checking.
6. Let others write about you too
Models trust information they see confirmed across independent sources. Industry directories, professional bodies, partner sites and press coverage all corroborate that you exist. For machine-readable company data see schema.org Organization.
Where not to spend your time
Do not pin your hopes on llms.txt. This file format presents site content to language models and has been widely discussed over the past year. The measured reality: adoption sits at roughly 10% across a study of 300,000 domains, and crawlers barely request it — out of more than 500 million AI bot visits over ninety days, only a few hundred targeted the file. Google has stated plainly that it does not support the format and does not plan to.
The file is harmless and takes ten minutes to add, and it has real value in narrow cases — for instance when a developer points a model directly at your documentation. Just do not treat it as your main visibility tool.
Do not mass-produce AI content. Ten variations of one topic advance neither in Google nor in model answers. One strong article beats five shallow ones.
Do not stuff keywords. That tactic stopped working in Google long ago and never worked in a language model.
How to measure it
- Manual checks. Once a month, put your fifteen key questions to ChatGPT, Perplexity and Gemini and record who gets named. Simple, but the most reliable method available.
- Server logs. Track
GPTBot,ClaudeBotandPerplexityBotvisits to see which pages are actually being read. - Referral sources.
chatgpt.comandperplexity.ainow appear separately in analytics. Volumes are small, but those visitors usually arrive close to a decision.
Conclusion
AI search is not a separate channel — it is a layer on top of the work you already do. Preparing for it takes no secret tool: ask the question clearly, answer it briefly and concretely, do not hide who you are, and do not block the crawlers.
Above all: a great deal of empty promise is sold in this field. Measure, distrust invented figures, and verify results on your own site. Get in touch if you would like us to check whether your site is ready.
Frequently asked questions
Are GEO and SEO separate disciplines?
No, they overlap heavily. A fast, technically sound site that covers its topic properly wins on both sides. GEO adds structured data, a question-and-answer format and concrete phrasing on top.
How soon do results show?
Question-answer blocks and structured data can show an effect within one or two months. The part that depends on brand recognition takes considerably longer, because model training data is refreshed only a few times a year.
Can a small company compete with the giants?
Hard on broad queries, entirely possible on narrow and local ones. Wikipedia and Reddit have no answer there, while your concrete experience fits exactly.
Will AI traffic replace Google traffic?
Not soon. Google is still far larger in volume. But visitors arriving from AI answers are usually better prepared, because the model has already moved them one step forward.
Do I need to rebuild my site?
Usually not. In most cases adding a question-and-answer section, structured data and author details to existing pages is enough.
